Field Calibration Development and Comparative Evaluation of Machine Learning and Physics-based Wind Estimation Methods for the Weather-sensing CopterSonde UAS

Abstract

The precise measurement of wind within the planetary boundary layer (PBL) is essential for enhancing weather forecasts, comprehending atmospheric phenomena, and facilitating various applications,. Traditionally, researchers have depended on radiosondes, tower-mounted instruments, and remote sensing technologies like Doppler lidar to get kinematic and thermodynamic characteristics of the lower atmosphere. While these methodologies yield significant data, they are frequently limited by cost, mobility, or temporal resolution, resulting in limitations in our knowledge of the PBL's intricate and swiftly changing circumstances. Uncrewed Aircraft System (UAS) present a viable option owing to their adaptability, cost-effectiveness, and capacity for precise and targeted in situ measurements. Nevertheless, affixing wind sensors to small drones may result in weight and aerodynamic drawbacks, in addition to being vulnerable to disturbances or flow distortion around the aircraft. The CopterSonde, a drone developed by meteorologists and engineers at the University of Oklahoma, has been designed and optimized for kinematic and thermodynamic measurements of the PBL. Unlike conventional multicopter platforms, the CopterSonde employs a model-based 3D wind estimation technique that relies on mathematical models. This method uses aerodynamic principles to infer wind vectors from the drone’s own response against wind and control inputs. Early phases of the presented study, building on the foundation of previous works, focused on refining parameters of the model-based methods by calibration through the use of Differential Evolution Optimizer (DEO), a cost function optimization technique. Calibration was performed using a combination of Doppler wind lidars, providing a robust reference for validation. These efforts highlighted the DEO method's ability to optimize the CopteSonde's dynamic model, yielding wind estimates with accuracy comparable to Doppler wind lidars. by minimizing the root mean square error (RMSE) between the ground truth and the CopterSonde’s estimations. The integration of a Linear Extended State Observer (LESO) into the model helped to reduce the influence of autopilot-induced perturbations and enhance the overall quality of the wind estimates. Extending upon this previous research, the presented work introduces advanced machine learning (ML) approaches to avoid reliance on complex mathematical models, offering a platform-agnostic and data-driven characterization of UAS for wind estimation. A selection process to determine an adequate machine learning method was described and implemented. After evaluating several candidate techniques based on predictive accuracy and computational efficiency, machine learning techniques were employed to analyze and learn patterns and relationships in an extensive pre-collected dataset that model-based methods fail to capture. ML methods may offer a better way to distinguish wind perturbations from coupled UAS dynamics. The resulting ML models not only refined the CopterSonde’s 3D wind estimation capabilities but also provided a new perspective on how to integrate data-driven methods with mathematical models. The combined efforts in this study present a comprehensive evaluation of various wind estimation techniques, from initial drag coefficient refinement using DE to advanced ML implementations. This holistic approach not only highlights the progress made in enhancing wind-sensing accuracy but also provides valuable insights and guidelines for future applications to avoid reliance on complex mathematical models, offering a platform-agnostic and data-driven characterization of UAS for wind estimations. Through a comparison of different methodologies, the work establishes a solid foundation for ongoing research in wind estimation and sets the stage for future advancements in UAS technology.